Towards Financial World Modeling
This presentation explores a systematic study of representation learning for financial markets, introducing Market-1T, a dataset with nearly one trillion observations of U.S. equities from 2008 to 2025. The work rigorously compares 18 encoder-training strategies across 32 market regimes, revealing that predictive performance and economically meaningful latent organization are fundamentally distinct properties. The findings establish critical infrastructure and evaluation protocols for building future financial world models that can support planning, counterfactual reasoning, and market-wide understanding.Script
Financial markets operate as high-dimensional, rapidly shifting ecosystems where thousands of assets interact across changing regimes. The researchers behind this work ask a deceptively simple question: what makes a useful representation of the financial world?
They built Market-1T, a dataset containing nearly one trillion observations at one-second resolution covering all U.S.-traded equities across 17 years. This scale and temporal span make it possible to evaluate representations across dozens of distinct market regimes rather than a single snapshot.
The study compares 18 encoder-training strategies, including supervised models trained on specific targets, 14 self-supervised configurations that vary both objective and augmentation, and several pretrained baselines. Each approach embeds market observations into a latent representation that can be tested for what information it preserves.
The central finding is striking: predictive performance and economically meaningful organization are nearly independent. Supervised models excel at forecasting 15-minute returns and volatility changes, but many fail to organize the latent space by firm identity, industry, or shared market factors. Meanwhile, some self-supervised methods preserve market structure but struggle to beat even a randomly initialized encoder at prediction.
Evaluation across 32 months spanning 17 years reveals that market regime explains over 97% of return-prediction variance. A model that looks strong in one period may falter in another, making single-period benchmarks deeply misleading for financial applications.
This work establishes the foundation for financial world models by showing that useful representations must preserve multiple dimensions: not just predictive power, but also the structure of firms, industries, and market-wide conditions. To explore the full dataset, evaluation protocols, and findings in depth, visit EmergentMind.com where you can create your own video summaries of cutting-edge research.
